代理自管的程序性记忆:skill_manage 工具可创建、修补、禁用、删除 agent 自己的技能(用户级/项目级双作用域),带删除守卫(agent 创建标记、pin、路径限制)与触发纪律提示。
安装
# npm 包(预构建)
dsh plugin --profile web add dsh-skill-manage
# GitHub 源码(首次需按提示配置 allowBuilds 构建授权后重试)
dsh plugin --profile web add github:fuxin123z/dsh-skill-manage
装任何插件都等于在你的机器上跑第三方代码,权限和你本人一样大——能读你的文件、用你的凭据、访问网络,工具审批管不到它。GitHub 来源的插件还会在安装时执行构建脚本——pnpm 默认拦截,所以安装可能停在 ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED 或 ERR_PNPM_IGNORED_BUILDS;dsh 会打印出需要添加的确切键名,把它加进该 profile 的 pnpm-workspace.yaml 的 allowBuilds 下,重跑一次即可装上。放行构建本身就是一次信任判断:请只安装可信来源,并尽量锁定 commit(github:owner/repo#sha)。
README
该插件的 README 只有英文版本。
A DSH (DeepSeek Harness) plugin that gives the agent procedural memory: a skill_manage tool for authoring its own skills — create, patch, disable, delete — with layered delete guards. Built on skill-filesystem's hot-reload watcher, so a skill created mid-session is usable in the same session, no restart.
Pair with dsh-auto-memory for declarative memory: that one remembers facts (logs, notes, preferences), this one remembers how to do things (workflows, pitfalls, procedures).
What it adds
skill_manage tool — actions: create / patch / edit / delete / disable / enable / pin / unpin / write_file / remove_file / list
- Two scopes —
scope: 'user'(default,~/.dsh/skills, all workspaces) orscope: 'project'(<projectRoot>/.dsh/skills, this workspace only). The project root is resolved exactly like the harness's own skill lookup: walk up from the session cwd to the nearest.git, falling back to the cwd itself.listshows both scopes in one table. - Both on-disk layouts — directory skills (
<root>/<name>/SKILL.md, whatcreatewrites) and single-file skills (<root>/<name>.md), matching the two layouts theskill-filesystemwatcher actually loads.listshows alayoutflag; single-file skills refuse supporting-file actions. - Disable/enable (reversible) — toggles
disable-model-invocationin SKILL.md frontmatter, the same key the harness's skill catalog filters on (isModelInvocable). Disabling hides a skill from the model's catalog without touching its content; enabling restores it. Preferdisableoverdeletefor seasonal or off-context skills. - Pin/unpin (reversible) — toggles the same
pinnedfrontmatter flag the delete guard reads. A pinned skill cannot be deleted byskill_manage(patch/edit still allowed); pin skills that must survive cleanups. - Trigger discipline (English, static system-prompt section, cache-stable): create a skill when a complex task succeeded (5+ tool calls), errors were overcome, a user-corrected approach proved itself, or the user asks to remember a procedure; patch immediately when a skill hits uncovered pitfalls.
- Delete guards:
- only skills carrying the
created_by: agentfrontmatter marker are deletable — marketplace/user skills are refused pinned: truefrontmatter blocks delete (patch/edit still allowed)- path confinement to the resolved skills root; symlinked skill directories refused
- name-drift guard: patch/edit cannot silently rename a skill
- only skills carrying the
- Validation: name regex + length, frontmatter requires
name+description, description ≤1024 chars, SKILL.md ≤100k chars, supporting files ≤1 MiB, supporting paths confined toreferences/ templates/ scripts/ assets/ - Atomic writes (temp + rename) so the watcher never sees a half-written SKILL.md
- CRLF-safe frontmatter parsing
Install
Option A — via dsh plugin (recommended; handles both steps below automatically):
dsh plugin --profile web add dsh-skill-manage
Option B — manual npm install, in your DSH profile dir (e.g. ~/.dsh/profiles/web):
# 1. add the dependency
npm install dsh-skill-manage # or: pnpm add dsh-skill-manage
# 2. register the bundle in package.json → dsh.profile.bundles:
# "dsh": { "profile": { "bundles": [ ..., "dsh-skill-manage" ] } }
# 3. restart dsh web
The plugin needs no configuration. Log line [dsh-skill-manage] ready confirms it loaded.
From source (local link) — for development:
# in your DSH profile dir (e.g. ~/.dsh/profiles/web)
# 1. package.json dependencies:
# "dsh-skill-manage": "link:/abs/path/to/dsh-skill-manage"
# 2. dsh.profile.bundles: append "dsh-skill-manage"
pnpm install
# 3. restart dsh web
Develop & test
node test.mjs # guard-level smoke tests (73 cases, sandboxed DSH_HOME)
node loadtest.mjs # host-shape contract + real round-trip (set DSH_HOME to sandbox it)
CI runs both on every push/PR (.github/workflows/ci.yml).
v0 known limitations
- Nested category dirs (
<root>/<category>/<name>/SKILL.md) are not listed or operable — the harness watcher does not load them either (it loads only<root>/<name>/SKILL.mdand<root>/<name>.md) - No YAML flow collections in frontmatter (
tags: [a, b]stays a raw string) - Project scope assumes one root per session (the session cwd's git root)
License
MIT
链接
同类插件
zhu1090093659/dsh-web#packages/dsh-skill-explorer★ 8440
技能中心:按来源分级浏览已加载的全部 skill,启用/禁用模型调用、创建新技能、删除进可恢复回收站。
GanyuanRan/Aegis★ 1323
面向编码 Agent 的软件工程方法包,提供基线优先规划、系统化调试、提示词卫生、完成前验证,以及修复/退役双轨跟踪技能。
superdesigndev/superdesign-skill★ 628
在 Superdesign 画布上做 UI 与营销图的设计技能:先读代码库拿上下文、抽取现有设计系统,再通过 Superdesign CLI 生成并迭代可分支的设计稿、流程页与可复用组件。
linhay/harmony-next.skills★ 360
为 DeepSeek Harness 提供 HarmonyOS NEXT 技能包、离线 API 参考及 DevEco、HDC 与模拟器自动化指南。
dhicoc/dsh-reverse-skill★ 213
完整 reverse-skill(85 个 SKILL.md)的 DeepSeek Harness 插件:逆向工程、授权渗透测试与安全研究的技能路由包。
sandbaseai/sandbase-skills★ 201
通过文件系统 Skill provider 将 88 个研究、社交情报、营销与商业 Agent Skills 挂载到 dsh。
社区评论
评论公开保存在 GitHub Discussions。加载评论会连接 GitHub 和 Giscus;发表内容需要 GitHub 账号。